Version 2.0 | Developed by BankGPT Team

BankGPT is a next-generation AI-powered personal loan processor that replaces static forms with a dynamic, conversational, multi-agent system. Instead of overwhelming users with long forms, BankGPT guides applicants through an intuitive chat workflow, performing verification, underwriting, fraud checks, and sanctioning in real time.

• Users interact naturally through chat instead of forms.
• Dynamic prompts appear only when needed (e.g., salary slip upload for high amounts).

A Master Agent coordinates the workflow across:

• Verification Agent
• Underwriting Agent
• Fraud Agent
• Sanction Agent

Automatically decides the journey:

• FAST_TRACK → Instant approval
• CONDITIONAL_REVIEW → Requires additional documents
• HARD_REJECTION → Stops with clear explanation

A transparency sidebar displays:

• Credit score insights
• Underwriting reasoning
• Fraud check explanations
• Agent-level status updates

At approval, BankGPT generates:

• PDF Sanction Letter (auto-download)

Conversation-driven workflow instead of linear forms.

User <--> Streamlit UI <--> Master Agent

Orchestration Layer:
    Phase 1: Sales
    Phase 2: Underwriting
    Phase 3: Conditional Review
    Phase 4: Sanction

Worker Agents:
    Verification Agent
    Underwriting Agent
    Fraud Agent
    Sanction Agent

Verification Agent <--> Mock CRM (JSON)
Sanction Agent --> PDF Generator

• Friendly onboarding
• Product pitch + fixed interest rate (11%)

Two-step verification:

1. Identity → Lookup via mock CRM
2. Eligibility → Compare request vs. pre-approved limit

Triggered if:

• Requested amount > pre-approved limit
• Score is decent (≥700)

Prompts:

• Salary slip upload
• Fraud scan

• Approval
• Auto-generated PDF sanction letter

git clone <repo_url>
cd loan_rag_streamlit
python -m venv venv

pip install -r requirements.txt

Note: For Voice & OCR Install FFmpeg and Tesseract on system path.

streamlit run app.py

python test_conversation_flow.py

loan_rag_streamlit/
├── app.py                     # Streamlit UI
├── master_agent.py            # Phase orchestration
├── session_manager.py         # Chat state persistence
├── agents.py                  # Verification, Fraud, Underwriting, Sanction agents
├── eligibility.py             # Fast-track / Hard-reject logic
├── test_conversation_flow.py  # Unit tests
├── data/
│   └── mock_db.json           # CRM mock database
├── outputs/                   # Generated sanction letters
└── requirements.txt

Mock CRM (excerpt from data/mock_db.json):

{
  "9876543210": {
    "name": "Amit Kumar",
    "credit_score": 780,
    "income": 65000,
    "blacklisted": false,
    "approved_amount": 1200000
  },
  "9998887776": {
    "name": "Neha Singh",
    "credit_score": 710,
    "income": 48000,
    "blacklisted": false,
    "approved_amount": 800000
  },
  "8887776665": {
    "name": "Ravi Sharma",
    "credit_score": 640,
    "income": 40000,
    "blacklisted": true,
    "approved_amount": 0
  }
}

Applicant: Amit Kumar Loan Request: ₹500,000 (<= ₹1.2M) Score: 780

• Verification: Pass
• Underwriting: Pass @ 10.5%
• Fraud: Clear
• Route: FAST_TRACK
• Approved
• Sanction PDF generated

Applicant: Neha Singh Loan Request: ₹1,200,000 (> ₹800k limit but < 2×) Score: 710

Flow:

• Conditional Review → Salary slip upload
• Fraud scan → Pass
• Approved → PDF generated

Applicant: Ravi Sharma Score: 640 Blacklisted: Yes

• Underwriting: Fail
• Fraud: Fail
• Route: HARD_REJECTION
• Rejected with XAI explanation

Score: 590 (<650)

• Immediate rejection
• Clear XAI reasoning

Unknown applicant

• Verification: Not found
• Route: Rejected → Manual Review

Input:

• Loan: ₹1,000,000
• Tenure: 10 yrs
• Rate: 11%
• Income: ₹100,000
• Existing EMI: ₹10,000

Output:

• EMI: ~₹13,218
• FOIR: 23.2%
• Band: Likely Eligible

Process:

1. Click microphone
2. Speak query
3. Auto-transcribed
4. Answered with RAG/Gemini

• Upload PNG/JPG
• OCR extracts text
• Auto-filled into conversation

Enter Applicant Info
      ↓
Click "Process Loan Application"
      ↓
Multi-Agent Pipeline:
    - Verification
    - Underwriting
    - Fraud Detection
    - Eligibility Routing
      ↓
FAST_TRACK  → Instant Approval → PDF
CONDITIONAL → Salary Slip Upload → Approval/Rejection

• Correct sales pitch
• Accurate routing (Fast/Conditional/Reject)
• Document upload appears only when needed
• XAI panel shows detailed reasoning
• Sanction PDFs generate properly